Machine learning-based identification of process deviations in deep drawing of a cylindrical cup

Uncertainty is an inherent feature of sheet metal forming processes. It arises from variations in material properties, tooling conditions and process parameters. These uncertainties can significantly impact the quality and consistency of formed components, yet they are often poorly characterized or entirely neglected in conventional workflows. Although widely used for design and optimization, Finite Element Analysis typically assumes deterministic input parameters, which limits its ability to reflect real-world process variability. This study presents data-driven approaches for identifying sources of uncertainty in deep drawing. Instead of treating uncertainty as an external input, this study uses supervised machine learning techniques to identify key uncertainty factors from observable forming outcomes such as thickness distribution, earing profile and punch force evolution. Two modeling strategies were explored: one based on individual single-output models, where each uncertainty source is predicted separately, and another based on a multi-output learning framework that simultaneously captures the interactions between multiple uncertainty factors. A simulation-derived dataset reflecting irregularities in geometric, material, and process factors was generated, including blank thickness, punch and die misalignments and geometric changes, friction, and force of the blank holder. The trained models were experimentally validated in specific setups and demonstrate the capability to infer the underlying uncertainty conditions based on the forming results. The results show that specific forming signatures can be traced back to distinct sources of uncertainty, providing a framework with potential application for offline process diagnosis, to be further explored for in-process deployment.